Nonnegative Tucker Decomposition

Key Points

• To develop a nonnegative tensor factorization framework based on the Tucker model with improved algorithms for multilinear data analysis.
• Formulated the Nonnegative Tucker Decomposition (NTD) model by applying nonnegativity constraints to the Tucker tensor framework.
• Developed multiplicative update algorithms, an efficient initialization procedure for faster convergence, and a sparseness control mechanism.
• Evaluated the decomposition method across several computer vision benchmark examples against existing NMF and NTF approaches.
• NTD demonstrated effective multilinear feature representation and extraction across computer vision tasks.
• The proposed initialization and sparseness control techniques improved computational convergence speed and interpretability compared to standard NMF and NTF methods.